New Zen microarchitecture details

[coercitiv]

How's 14ff an immature process if Samsung and Apple have been shipping chips with it in high volume for over a year now?

Since you mention Apple, we had 14nm going versus 16nm in A9 chips and it showed Samsung's process to use more power at top chip frequency.

We can also ask the same question from another point of view: how is 16nm a mature process when Nvidia still struggles to make enough GP104 chips to meet demand?

Let's take this a bit further, what did Intel have trouble with recently: chip design for Broadwell and Skylake... or making their 14nm process work? You know Intel, the company with the big fat R&D budget and flush with talent.

This is why people corroborate data from the Polaris launch and assume they might have problems at the fab. Not because their engineers can do no wrong.

 

[senseamp]

Since you mention Apple, we had 14nm going versus 16nm in A9 chips and it showed Samsung's process to use more power at top chip frequency.

We can also ask the same question from another point of view: how is 16nm a mature process when Nvidia still struggles to make enough GP104 chips to meet demand?

Let's take this a bit further, what did Intel have trouble with recently: chip design for Broadwell and Skylake... or making their 14nm process work? You know Intel, the company with the big fat R&D budget and flush with talent.

This is why people corroborate data from the Polaris launch and assume they might have problems at the fab. Not because their engineers can do no wrong.

Fine, just don't be surprised if/when Zen is also having "problems at the fab."
AMD is the one bragging about how great FinFet is:

 

[coercitiv]

Fine, just don't be surprised if/when Zen is also having "problems at the fab."
AMD is the one bragging about how great FinFet is:

Again, Intel was also bragging about how great 14nm Broadwell products were versus 22nm Haswell, yet as it turns out they had yield problems. Broadwell was also having issues clocking high, with most of the power advantage over 22nm disappearing into thin air as chips went past 3Ghz+ (mobile parts). It took them quite a while to get the process ironed out and only with Skylake and Kaby Lake are we seeing a return to steady increasing clocks in mobile.

Whatever AMD claims right now is no longer of interest, since we've seen how the Polaris launch went: the good news is yields are good enough to make it a real launch, the bad news is they had to accept all kinds of chip quality from the fab to get that kind of stock. The situation is best described as Meh... :sneaky:

Now, if what you're saying is that we should not default to blaming the fab if Zen comes out in poor shape, then I agree. But right now people are "working" with what little info they have, and looking at the recent GPU launch, it seems to me that the best quality chips set a line for how well AMD engineered them (top freq, power usage etc), while the low quality chips set the line for how well GF is doing their job.

 

[The Stilt]

Polaris 10 is probably the saddest GPU I've ever seen from ATI / AMD

Based on few samples the GPU basically stops scaling at ~1350MHz region. Beyond that point the voltage must be increased so heavily that the power draw of the GPU get's completely out of control. The GPU appears to be also almost immune to improved cooling and the gains from lower temperatures are almost non-existent, as long as you can prevent the GPU from overheating of course. The 28nm processes from GlobalFoundries behave quite similarily on AMD APUs.

244W power draw (DCR, ASIC only) was recorded during 3DMark FS at 1.3000V / 1500MHz frequency (water cooled). A pretty nasty increase, considering that the GPU is rated for just 110W TDP at stock. At stock the power consumption was around 123W and many of the cards appear to slightly exceed their power budget at stock.

 

[dark zero]

Polaris 10 is probably the saddest GPU I've ever seen from ATI / AMD

Based on few samples the GPU basically stops scaling at ~1350MHz region. Beyond that point the voltage must be increased so heavily that the power draw of the GPU get's completely out of control. The GPU appears to be also almost immune to improved cooling and the gains from lower temperatures are almost non-existent, as long as you can prevent the GPU from overheating of course. The 28nm processes from GlobalFoundries behave quite similarily on AMD APUs.

244W power draw (DCR, ASIC only) was recorded during 3DMark FS at 1.3000V / 1500MHz frequency (water cooled). A pretty nasty increase, considering that the GPU is rated for just 110W TDP at stock. At stock the power consumption was around 123W and many of the cards appear to slightly exceed their power budget at stock.

I would be worried abour Pascal... higher clocks... with less effective results than expected. Only they won since they went with big chips at first... but I feel that GP 100 won't deliver the promised performance they sell.

 

[superstition]

Polaris 10 is probably the saddest GPU I've ever seen from ATI / AMD

Based on few samples the GPU basically stops scaling at ~1350MHz region. Beyond that point the voltage must be increased so heavily that the power draw of the GPU get's completely out of control. The GPU appears to be also almost immune to improved cooling and the gains from lower temperatures are almost non-existent, as long as you can prevent the GPU from overheating of course. The 28nm processes from GlobalFoundries behave quite similarily on AMD APUs.

244W power draw (DCR, ASIC only) was recorded during 3DMark FS at 1.3000V / 1500MHz frequency (water cooled). A pretty nasty increase, considering that the GPU is rated for just 110W TDP at stock. At stock the power consumption was around 123W and many of the cards appear to slightly exceed their power budget at stock.

So, it's the saddest for overclocking. Since this is a Zen CPU thread, though, I'll ask how sad its overclocking is in comparison with locked Intel CPUs.

Both, after all, are aimed at the mid-priced mainstream buyer.

 

[frozentundra123456]

So, it's the saddest for overclocking. Since this is a Zen CPU thread, though, I'll ask how sad its overclocking is in comparison with locked Intel CPUs.

Both, after all, are aimed at the mid-priced mainstream buyer.

If Zen is only able to compete with locked Intel i5 chips, it is the mother of all failures.

 

[superstition]

If Zen is only able to compete with locked Intel i5 chips, it is the mother of all failures.

I guess what I was trying to get at is that we're talking about a $200 product. Polaris 10 isn't supposed to be a high-end offering. As such, its overclocking prowess is more of an academic angle than a central one — since it apparently delivers adequate performance at its price point, provided its power delivery is suitably regularized.

While it would be nice, of course, for customers to be able to unlock quite a bit more additional performance with minimal extra cost — is it actually true that the 480 doesn't deliver adequate performance at its price point to be competitive with Nvidia's offerings?

If Nvidia offers a product of equal stock performance at the same price point but which can be overclocked quite a lot without high power consumption then, yes, I can see the 480 being disappointing. However, AMD may be able to respond by reducing price further or something.

If the foundry's lack of quality product is the main issue with the 480 then we can hope that the delay between Polaris 10 and Zen will give GF/AMD enough time to improve things significantly. The Stilt mentioned, for instance, the improvement made to 32nm SOI. Design rules proliferate as nodes shrink, though, so there is only so much — in terms of development resources — to go around.

 

[IntelUser2000]

Price is the most elastic part of the whole equation.

If it misses design targets, they can lower price. Of course they can choose not to, but no doubt AMD knew about Polaris issues already and wouldn't have been competitive at all at higher prices.

The chances are GTX 1060 is going to spoil whatever AMD has for their RX 480. At least for its lower quality it'll be somewhat cheaper.

However, AMD may be able to respond by reducing price further or something.

Reducing official MSRPs may be a bit too much for them. But they can do subtle changes. Like offer the 4GB models with 8GT/s memory, rather than 7GT/s, and make it more widely available. Then "Polaris 10 @ $199!" would come true at least.

 

[The Stilt]

So, it's the saddest for overclocking. Since this is a Zen CPU thread, though, I'll ask how sad its overclocking is in comparison with locked Intel CPUs.

Both, after all, are aimed at the mid-priced mainstream buyer.

No, the saddest GPU like I said.

AMD claimed up to 70% improved PPW from 14nm FinFet alone and an additional 110% (up to) from design / architectural improvements. Yet the actual, average PPW is barely 10% higher than on their previous generation 28nm GPU (R9 Fury X)*
*TPU 1080p - 2160p

On average the brand new Polaris 10 is 12% behind the PPW of GTX 980 made on 28nm process and released nearly two years ago. On average Polaris 10 provides > 44% lower PPW than nVidia's Pascal based GTX 1080.

The saddest thing is that nVidia does all this on a GPU, which is > 35% larger and clocked 50% higher (1266 vs. 1898MHz).

Prior Polaris and Pascal AMD was even remotely competitive with GPUs. Now the difference is similar what we have used to see between AMD and Intel CPUs. At this rate AMD won't see the end of the decade, unfortunately.

 

[24601]

No, the saddest GPU like I said.

AMD claimed up to 70% improved PPW from 14nm FinFet alone and an additional 110% (up to) from design / architectural improvements. Yet the actual, average PPW is barely 10% higher than on their previous generation 28nm GPU (R9 Fury X)*
*TPU 1080p - 2160p

On average the brand new Polaris 10 is 12% behind the PPW of GTX 980 made on 28nm process and released nearly two years ago. On average Polaris 10 provides > 44% lower PPW than nVidia's Pascal based GTX 1080.

The saddest thing is that nVidia does all this on a GPU, which is > 35% larger and clocked 50% higher (1266 vs. 1898MHz).

Prior Polaris and Pascal AMD was even remotely competitive with GPUs. Now the difference is similar what we have used to see between AMD and Intel CPUs. At this rate AMD won't see the end of the decade, unfortunately.

The writing has been on the wall since at least Tahiti, as AMD doesn't seem to have any new GPU architectures planned after GCN.

The AMD board is simply doing what it always does, claw up as much benefit for themselves as possible before donning their golden parachutes.

We saw this play out basically since AMD's inception. The board cares about no-one but themselves. Shareholders, employees, partners, consumers be damned.

 

[superstition]

AMD claimed

Skylake was also supposed to be revolutionary. I even fell for the hype and advised people to wait on purchases. Never again. Broadwell C outperforming it with a lower clock rate and power consumption was just too funny.

I share Peter Bright of ArsTechnica's disappointment with Skylake — that it didn't have even one SKU with EDRAM. Furthermore, I think it's scandalous to offer Broadwell E after all the Skylake hype and after putting Skylake on the desktop. All the stuff about how Intel always does that doesn't change that in my view. It's just the Appeal to Tradition fallacy.

On average the brand new Polaris 10 is 12% behind the PPW of GTX 980 made on 28nm process and released nearly two years ago.

It's PPD where the 980 was never much of a king, though, eh? I agree that, from a technical standpoint, it was quite a nice product. But, I wonder how much its tremendous performance per watt kept up with the heavy overclocking most people did with the part.

AMD could have improved the PPW of Polaris 10 by keeping stock clocks down but it seemed to be targeting the 970 and, from the little I've read, done a good enough job to replace it as the mainstream market leader.

Perhaps just enough is enough for the time being?

On average Polaris 10 provides > 44% lower PPW than nVidia's Pascal based GTX 1080.

But again, PPD.

The saddest thing is that nVidia does all this on a GPU, which is > 35% larger and clocked 50% higher (1266 vs. 1898MHz).

Maybe. Or, it could be the price tag that Nvidia always charges — since we are talking about a product targeting the $200 bracket.

 

[The Stilt]

Skylake was also supposed to be revolutionary. I even fell for the hype and advised people to wait on purchases. Never again. Broadwell C outperforming it with a lower clock rate and power consumption was just too funny.

I share Peter Bright of ArsTechnica's disappointment with Skylake — that it didn't have even one SKU with EDRAM. Furthermore, I think it's scandalous to offer Broadwell E after all the Skylake hype and after putting Skylake on the desktop. All the stuff about how Intel always does that doesn't change that in my view. It's just the Appeal to Tradition fallacy.

It's PPD where the 980 was never much of a king, though, eh? I agree that, from a technical standpoint, it was quite a nice product. But, I wonder how much its tremendous performance per watt kept up with the heavy overclocking most people did with the part.

AMD could have improved the PPW of Polaris 10 by keeping stock clocks down but it seemed to be targeting the 970 and, from the little I've read, done a good enough job to replace it as the mainstream market leader.

Perhaps just enough is enough for the time being?

But again, PPD.

Maybe. Or, it could be the price tag that Nvidia always charges — since we are talking about a product targeting the $200 bracket.

So these facts become irrelevant because: A) Intel has made overly optimistic statements regarding Skylake, B) PPD.

What's next, because of PPI (performance per inch)? :sneaky:

If PPD mattered that much and both the absolute performance and the PPW would be secondary, AMD CPUs and APUs would be selling like hotcakes.

And mind you that AMD claimed 2.8x PPW for RX 480 (as a product and not as a generic improvements for the architecture & manufacturing process). So even if it was clocked outside it's optimal operating frequency (which it definitely is), it should do "slightly" better than the ~10% higher PPW.

 

[superstition]

So these facts become irrelevant because: A) Intel has made overly optimistic statements reganding Skylake, B) PPD.

What's next, because of PPI (performance per inch)? :sneaky:

If PPD mattered that much and both the absolute performance and the PPW would be secondary, AMD CPUs and APUs would be selling like hotcakes.

You're really saying the GPU market hasn't seen mid-priced products dominate sales?

 

[IntelUser2000]

I agree in general with The Stilt. Price is very elastic, everything else is very fixed.

They could have charged more if it performed much better. Initial pricing of some FX CPUs demonstrated AMD doesn't necessarily have the best grasp of PPD.

Also, the more trusted/known company naturally can charge more for their products. It takes years and years of leadership to change that.

Skylake was also supposed to be revolutionary. I even fell for the hype and advised people to wait on purchases. Never again. Broadwell C outperforming it with a lower clock rate and power consumption was just too funny.

I'd reckon such technical "failures" are nothing compared to what's coming. It'll only get harder and complex. We thought Nvidia failed with Pascal, then Polaris released. If those best at their game fails, those that are less than best won't do better.

I share Peter Bright of ArsTechnica's disappointment with Skylake — that it didn't have even one SKU with EDRAM.

It did not make sense at all, for Intel to release Skylake Desktop part with eDRAM so shortly after Broadwell with eDRAM. They rarely ever cannibalize their products. Skylake eDRAM 3 months after Broadwell eDRAM is the very definition of that. I knew something was up. The reality was they were simply not ready. See how they just released BGA Iris Pro Skylake.

 

[AtenRa]

And mind you that AMD claimed 2.8x PPW for RX 480 (as a product and not as a generic improvements for the architecture & manufacturing process).

According to AMD, 2.8X is between RX470 110W TDP vs R9 270X 180W TDP in 3D Mark Fire Stike.

Edit: What is more interesting is the Mobile RX480M (Polaris 11 16x CUs) vs R9 280X (14x CUs).
According to AMD, in 3D Mark Fire strike the RX480M got a score of 4070 with a Board Power of 35W when R9 M280X managed a score of 3500 with a Board Power of 82W (Cat 16.20).
That is also ~2.8x Perf/Watt

 

[coercitiv]

According to AMD, 2.8X is between RX480 150W TDP vs R9 290 275W TDP in 5 games at 1440p (Ashes, GTAV, Witcher 3, Project Cars and Assassins Creed Syndicate).

Edit: What is more interesting is the Mobile RX480M (Polaris 11 16x CUs) vs R9 280X (14x CUs).
According to AMD, in 3D Mark Fire strike the RX480M got a score of 4070 with a Board Power of 35W when R9 M280X managed a score of 3500 with a Board Power of 82W (Cat 16.20).
That is also ~2.8x Perf/Watt

Really?! I didn't know RX 480 is 50% faster than R9 290.

 

[AtenRa]

Really?! I didn't know RX 480 is 50% faster than R9 290.

My mistake I was looking at the wrong notes, fixed , it is RX 470 vs R9 270X.

I dont know if they have made same testing with RX 480.

 

[majord]

Comparing perf/watt is tricky with only one product to gauge off - Just look how wildy PPW varies in those TPU charts on products with the same process and architecture - The different core configs, clockspeeds, ram amounts, and HBM (in the case of fury line) all play abig roll so it's not easy to really see where Polaris truely fits until we see RX 470 and so on.

https://www.techpowerup.com/reviews/ASUS/R9_380X_Strix/24.html

If you look at the 380X (which isn't included in the PPW spread for the 480 review) , you can see the RX 480 has 50% higher PPW - which is in line with Nvida's 1080 and 1070, and what they achieved against their similarly configured predecessors.

adding RX 470 and GTX960 to the mix will paint a better picture I think.

 

[KTE]

No, the saddest GPU like I said.

AMD claimed up to 70% improved PPW from 14nm FinFet alone and an additional 110% (up to) from design / architectural improvements. Yet the actual, average PPW is barely 10% higher than on their previous generation 28nm GPU (R9 Fury X)*
*TPU 1080p - 2160p

On average the brand new Polaris 10 is 12% behind the PPW of GTX 980 made on 28nm process and released nearly two years ago. On average Polaris 10 provides > 44% lower PPW than nVidia's Pascal based GTX 1080.

The saddest thing is that nVidia does all this on a GPU, which is > 35% larger and clocked 50% higher (1266 vs. 1898MHz).

Prior Polaris and Pascal AMD was even remotely competitive with GPUs. Now the difference is similar what we have used to see between AMD and Intel CPUs. At this rate AMD won't see the end of the decade, unfortunately.

If the process has problems, like AMDs 65nm, AMD always fudges the performance figures to be best case in a one-off corner-case benchmark. So the general figure ends up VERY different to their claims.

As with Polaris.

From what I'm seeing, 14nm is their 65nm all over again.

Or G80 vs R600.

I did say Phenom, right from the beginning... I don't see how Zen can possibly be competitive with this process.

 

[majord]

I have no doubt 14nm has issues, but disagree it's 65nm all over again..

65nm brought quite minimal power consumption improvement, and a frequency regression.

14nmLPP vs 28nm SHP is at least a substantial improvement in both areas. This is quite obvious i think

 

[StrangerGuy]

The writing has been on the wall since at least Tahiti, as AMD doesn't seem to have any new GPU architectures planned after GCN.

The AMD board is simply doing what it always does, claw up as much benefit for themselves as possible before donning their golden parachutes.

We saw this play out basically since AMD's inception. The board cares about no-one but themselves. Shareholders, employees, partners, consumers be damned.

I'm getting Phenom I and R600 flashbacks with 2016's AMD all over again, plus I could care less or wait for AMD's future "what-if" products with their lousy track record when I can get a GTX 1070 now and having a still excellent CPU 2 years ago.

If they are hoping Intel and NV are going to screw up like Netburst or GT200/Big Fermi as part of their grand "strategy", more power to them. Just don't be surprised if they won't even see the end of 2017.

 

[Arachnotronic]

If they are hoping Intel and NV are going to screw up like Netburst or GT200/Big Fermi as part of their grand "strategy", more power to them. Just don't be surprised if they won't even see the end of 2017.

I think AMD sees the end of 2017, but I think those who are looking for AMD to return to its former glory will be disappointed. They just can't afford to spend what Intel and NVIDIA spend individually in their respective markets, let alone as much as both.

 

[Arachnotronic]

If the process has problems, like AMDs 65nm, AMD always fudges the performance figures to be best case in a one-off corner-case benchmark. So the general figure ends up VERY different to their claims.

As with Polaris.

From what I'm seeing, 14nm is their 65nm all over again.

Or G80 vs R600.

I did say Phenom, right from the beginning... I don't see how Zen can possibly be competitive with this process.

Why is it the process's fault? Why is it never the fault of the company that actually designed the product?

Anyway, Polaris got a big jump in perf/watt over Tonga, no doubt process related. But AMD's architecture was so behind NVIDIA's to begin with that expecting to hit the perf/watt levels that NVIDIA did with Pascal on 16FF+ (which is a superior process, but not by THAT much) was just unrealistic wishful thinking.

 

[AtenRa]

I will suggest to revisit the perf/watt again with RX 470 vs GTX 1060 (Reference $249 cards) in DX-12 games throughout the year. Then we can talk IF and how much behind is GCN 1.4 architecture and the 14nm LPP process.